fix(chat): stop offering trust tiers a spawn approval cannot grant (#5400)

[bolichen97]

Summary

Closes #5400.

On a spawn approval in the chat activity viewer, the trust dropdown offered "Trust this command" / "Trust all <base> commands" / "Trust all tools", and picking any made the card report "Trusted" — but no standing trust was ever granted. Spawn approvals resolve through the one-shot api.resolveApproval endpoint, which has no trust verb, so every tier collapsed to a plain approve. The UI overstated the grant: the next identical call prompts again, and the user cannot tell whether trust "did not stick" or was silently downgraded. This is #5203's consent-integrity defect in the opposite direction, and the fix takes the same narrow-the-surface direction #5202 took for the channel surface.

What changed

• website/src/pages/chat/ActivityViewer.tsx — the spawn-approval card renders only Approve / Reject (the TrustDropdown is gone), and decisionLabel can only ever read Approved / Rejected. The approval_type === 'chat' branch in ApprovalEntry.onAction (and its slot prop / pattern param) is deleted: ApprovalEntry is module-local with a single mount behind the isSpawnApproval filter, so the branch was unreachable — and had it ever become reachable it would have produced the inverse mismatch (silently withholding tiers the chat surface genuinely grants). The real chat-approval trust path (ChatInput → approveChatSlot) is untouched.
• i18n catalogs — the three now-unreferenced keys (pages.chat.activityViewer.trusted / trusted_base / trusted_command) are removed from all 13 catalogs (the deadKeys gate is a ratchet); en-XA regenerated via npm run i18n:pseudo.
• website/src/test/ActivityViewerCoverage.test.tsx — replaces the trust-dropdown test with one pinning the honest set: the action row is exactly Approve/Reject (asserted by length and order) and the resolved card reads "Approved", never "Trusted". Mutation-verified: the new test fails (and is the only failure) with the fix reverted. The now-unneeded Radix dropdown vi.mock is removed — verified nothing else in the file renders a Radix dropdown (40/40 pass without it).
• Capture harness — website/capture/spawn-approval-trust.{html,tsx} mounts the REAL ActivityViewer with a pending spawn approval through its own toolLog seam; website/scripts/capture-spawn-approval-trust.mjs asserts each frame's action row before writing (the after-assertion fails on pre-fix code, so an after-frame cannot be shot from the old card by mistake).

Screenshots (SHA-pinned)

	Before	After
Dark
Light

Tests

• npx tsc -b clean; npx vitest run targeted: ActivityViewer suites (75), full src/i18n (629), TrustDropdown — all green; full suite runs in CI.
• i18n gates: check-i18n-keys (12674 refs resolve), lint:i18n at baseline, deadKeys ratchet intact (keys and references removed together).
• eslint: 0 errors, no warnings in changed files; jscpd clean.
• Pre-push review: Opus (claude-opus-5) — no blocking findings, 2 fixable advisories addressed in this commit (dead-branch deletion, exact-match capture wait); a third (latent same-class wiring in CollapsibleToolGroup, unreachable today) is being filed as a follow-up issue. GPT lane degraded (gpt-5.6-sol and gpt-5.6-terra both stalled at 30 min); backstopped by a contract-driven self-review against codex-review.yml's WHAT-BLOCKS list — the server-side GPT review lane on this PR remains the authoritative gate.

[github-actions]

UX Review (Fable 5) — ✅ PASS

UX-level review of a8aa9b23b0352a8ba1fd113b00490f08d296659c — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: PASS

Removing the trust tiers that never granted trust makes the card honest — Approve/Reject now says exactly what happens, in both themes.

[UX-REVIEWED] a8aa9b2

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of a8aa9b23b0352a8ba1fd113b00490f08d296659c — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

The diff checks out against the description: isSpawnApproval at ActivityViewer.tsx:244 excludes approval_type === 'chat', confirming the deleted chat branch was unreachable; TrustDropdown remains in use by the real chat trust path (ChatInput.tsx, ApprovalCard.tsx), which is untouched. The capture harness and screenshot directories follow long-standing repo convention. Removing the false affordance (rather than building trust semantics into the one-shot resolveApproval endpoint) is the narrower, reversible option and matches the direction of the cited precedents.

Design-Verdict: PASS

Removing an affordance the backend never honored fixes the consent-integrity mismatch at its root; the narrower surface beats retrofitting trust verbs onto a one-shot endpoint.

[DESIGN-REVIEWED] a8aa9b2

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of a8aa9b23b0352a8ba1fd113b00490f08d296659c and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] a8aa9b2

False positive or not applicable? A repository writer can comment:
/ai-review override gpt a8aa9b23b0352a8ba1fd113b00490f08d296659c: <one-sentence reason>

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed a8aa9b23b0352a8ba1fd113b00490f08d296659c — this comment is updated in place on each push.

Review details

The diff is a clean, well-scoped removal. The isSpawnApproval filter (line 244) excludes approval_type === 'chat', so the removed chat branch in onAction was dead for this card; the resolveApproval path is preserved. Grep confirms no surviving references to the removed activityViewer.trusted* keys. decisionLabel now carries only approved/rejected, the only values onAction can set. Capture files are test-only tooling, not bundled.

No findings.

[OPUS-REVIEWED] a8aa9b2

Verdict parsed from the review's SHA-scoped output markers for commit a8aa9b23b0352a8ba1fd113b00490f08d296659c.

False positive or not applicable? A repository writer can comment:
/ai-review override fable a8aa9b23b0352a8ba1fd113b00490f08d296659c: <one-sentence reason>

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of a8aa9b23b0352a8ba1fd113b00490f08d296659c — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Verification is complete. The fix checks out: the spawn card resolves through api.resolveApproval (website/src/api/client.ts:2459, approve|reject only), the deleted chat branch was unreachable (isSpawnApproval filters approval_type !== 'chat', single mount), the removed i18n keys have zero remaining references, and the capture/screenshot files follow the documented temp-screenshots/README.md convention with 225 sibling capture-*.mjs scripts. I counted one latent same-class sibling (CollapsibleToolGroup's trust→approve mapping in ChatPage), which is unreachable today and declared as a follow-up.

First-Principles-Verdict: PASS

The card claimed a standing trust grant the one-shot endpoint cannot make; deleting the claim is the subtractive fix at cause level.

What this change ships

Intent: stop the spawn-approval card from reporting "Trusted" when no trust is ever granted (#5400) — a FIX.

1. Spawn-approval card offers only Approve / Reject; Trust dropdown gone — justified (the fix)
2. Resolved spawn card can only read Approved / Rejected, never "Trusted" — justified (the fix)
3. Dead chat-approval branch and slot prop deleted from the card — rides along, declared, pure deletion
4. Three "Trusted…" strings removed from all 13 language catalogs — justified (deadKeys ratchet requires keys and references leave together)
5. Trust-dropdown test replaced by one pinning the Approve/Reject-only row — declared
6. New capture page + assert-before-shoot screenshot script — rides along, declared; matches the repo's 225 existing capture-*.mjs scripts
7. Four before/after PNGs under temp-screenshots/ — declared, sanctioned by temp-screenshots/README.md and its cleanup workflow

Watch

One counted unfixed sibling of the root cause (grepped toApiDecision|trust): ChatPage.tsx:4464 maps 'trust' → plain 'approve' behind CollapsibleToolGroup's Trust button — unreachable today (both mounts pass hasPermission={false}, ChatPage.tsx:6947, 6995) and declared as a filed follow-up. Note the general fix is not deleting the button: the third mount (app-sdk/ChatEmbed.tsx:183) routes trust verbatim through the slot endpoint and is honest.

Subtractions

Drop 'trust' from toApiDecision in ChatPage.tsx:4463-4464 — the only control that could send it is unreachable (both mounts pass hasPermission={false}), so the branch is dead wiring of exactly the class this PR removes.

[FIRST-PRINCIPLES-REVIEWED] a8aa9b2